Home / Research / AI Model Deprecation Statistics

Market research · Published October 8, 2026

AI Model Deprecation Statistics: 2026 Report

Companies that sell AI models through an API retire older versions on a schedule, and any app still calling a retired model gets errors until its developers move it to a newer one. We studied AI model deprecation across six platforms (the OpenAI API, Anthropic's Claude API, Google's Gemini API and Vertex AI, Mistral's API and Amazon Bedrock) to measure how much warning each one gives before it shuts a model off, how long models stay available after release, and how notice periods have changed since 2023.

We collected the dates from each platform's deprecation or model lifecycle page on October 8, 2026. Those pages list 283 model shutdowns, 213 that had already happened and 70 scheduled for later dates. The notice period is the number of days from the date a vendor announced a model's deprecation to its shutdown date. OpenAI, Anthropic and Mistral list both dates. For Amazon Bedrock we use the date a model entered what AWS calls the Legacy state, which is when AWS tells customers the end-of-life date. Google's two pages list release and shutdown dates without announcement dates, so Google models are not part of the notice figures.

Lifespan is the number of days from a model's release to its shutdown. Google lists a release date for most models. OpenAI and Anthropic do not, so for their models we start from the date written into the model's snapshot name, such as March 7, 2024 for claude-3-haiku-20240307, and leave out models whose names carry no full date. Mistral lists no release dates, so its models are not part of the lifespan figures. Shutdowns scheduled after October 8, 2026 count at their announced date.

We left out API endpoints, features such as OpenAI's Assistants API, fine-tuned models, cutoffs for new fine-tuning jobs, and Google's managed agents. When a model appears in more than one OpenAI notice, we use the most recent one. Cohere's deprecation page gives one effective date per notice and no separate shutdown date for most models, so Cohere is not included.

Stated Minimum Notice Before an AI Model Is Retired

Minimum Retirement Notice Stated by Each Platform, as of October 8, 2026

• OpenAI and Amazon Bedrock both set six months as the standard notice for regular models. Anthropic's stated minimum of 60 days is about one-third as long.
• Anthropic's page says its dates apply to the Claude API, Claude Platform on AWS and Microsoft Foundry, and that Amazon Bedrock and Google Cloud set their own retirement schedules for Claude models.

AI Model Deprecation Notice Periods by Vendor

Days From Deprecation Notice to Shutdown, by Platform

• OpenAI's median notice was 182 days, close to six months. Three in four OpenAI models got 180 days or more.
• Each of the 12 Amazon Bedrock models spent 181 to 185 days in the Legacy state before its end-of-life date.
• Anthropic's median notice was 63 days, and Mistral's was 70 days.
• Mistral gave the shortest notice of any platform. Its table lists June 6, 2025 as both the deprecation date and the retirement date for Codestral Mamba (open-codestral-mamba).
• OpenAI's shortest notices were 3 days for three Codex models (code-davinci-001, code-cushman-001 and code-cushman-002), announced March 20, 2023 and shut down March 23, 2023, and 20 days for gpt-5.4-cyber, announced September 11, 2026 and removed October 1, 2026.
• The longest notice was OpenAI's 458 days for gpt-3.5-turbo-0301, from June 13, 2023 to September 13, 2024.

How AI Model Deprecation Notice Has Changed Since 2023

Median Deprecation Notice by Year Announced

• Anthropic's four deprecation notices in 2026 gave 60, 62, 61 and 61 days, all within two days of its 60-day minimum. In 2025 its notices ranged from 62 days for Claude Haiku 3.5 to 189 days for Claude Opus 3.
• Mistral's three notices from late 2025 each gave 30 days or less. Notices dated October 31, November 6 and December 2, 2025 covered 13 models and gave 30, 24 and 29 days.
• Mistral's 2026 median rose to 91.5 days, but its September 29, 2026 notice retired labs-leanstral-1-5 and mistral-ocr-4-0 the next day, September 30, 2026.
• Five of OpenAI's nine notices in 2026 gave 182 to 184 days for every model they covered. Its April 22, 2026 notice covered 25 models: 11 older snapshots such as gpt-4o-2024-05-13 got 184 days, and 14 chat, Codex, deep research, search preview and mini audio models got 92 days. Its October 1, 2026 notice gave three GPT models 182 days and four text-to-speech models 97 days.

AI Model Lifespan From Release to Shutdown

Days From Release to Shutdown, Retired and Scheduled Models

• Preview models on Google's Gemini API had a median lifespan of 217 days, and 28 of 33 were shut down or scheduled to shut down within 12 months of release. The shortest was gemini-3.1-flash-lite-preview, released March 3, 2026 and shut down May 25, 2026, 83 days later.
• Stable Gemini API models had a median lifespan of 433 days, about twice that of preview models.
• Vertex AI retired gemini-1.5-pro-001, gemini-1.5-pro-002, gemini-1.5-flash-001 and gemini-1.5-flash-002 exactly 365 days after release.
• The two shortest Vertex AI lifespans belong to newer models that Vertex lists in its table for shorter availability periods. Vertex's gemini-3.6-flash, released July 21, 2026, retires November 19, 2026, after 121 days, and gemini-3.7-flash retires after 168 days.
• The four longest lifespans belong to embedding models. Vertex AI's multimodalembedding@001, released February 12, 2024, retires April 1, 2027, 1,144 days later.
• Claude Opus 4.1 and Claude Sonnet 3.7 were retired 365 days after the dates in their model names. Claude Haiku 3 lasted 774 days, the longest of the Anthropic models.
• OpenAI's gpt-4o-2024-05-13 is scheduled to shut down October 23, 2026, 893 days after its snapshot date, the longest lifespan among OpenAI models with a dated name.

AI Model Deprecation Statistics by Year

Model Shutdowns by Year, by Platform

• The six platforms shut down 99 models from January 1 to October 8, 2026, more than in all of 2025. Another 70 shutdowns are scheduled for later dates.
• OpenAI shut down 29 of its 34 models retired in 2024 on a single day, January 4, 2024. That day it retired older GPT-3 and GPT-3.5 completion models, its edit models and its first-generation text embedding models.
• The current Gemini API page lists no shutdowns before 2025, and the Amazon Bedrock page lists none before 2026, so the early years undercount those two platforms.

Same Model, Different Shutdown Dates on Cloud Platforms

Shutdown Dates for the Same Model on Two Platforms

• Anthropic and Amazon Bedrock both list shutdown dates for three Claude models, and each stays on Bedrock 121 to 156 days longer than on Anthropic's own API.
• Of 18 OpenAI model snapshots with a shutdown date on both OpenAI's API and Microsoft's Azure OpenAI schedule, 13 stay available longer on Azure and 5 shut down earlier there. The median gap is 27 days later on Azure.
• Five of the six models that Google lists with dates on both the Gemini API and Vertex AI share the same shutdown date. The exception is text-embedding-004, which Vertex AI keeps 442 days longer.

Limits of the AI Model Deprecation Data

Each platform's page covers a different stretch of time. OpenAI's lists model shutdowns back to 2023, while Google's Gemini API and Amazon Bedrock pages list only recent shutdowns, so year-over-year counts compare platforms with different histories. Notice periods are measured per model, and one notice can cover many models, such as OpenAI's July 6, 2023 notice for 29 models. Lifespans for OpenAI and Anthropic start from the snapshot date in the model name, which may differ from the public launch date. Shutdown dates after October 8, 2026 are the dates announced as of that day; Google's Vertex AI page says its retirement dates may be extended but will not move earlier.